Aleph Alpha releases 78-billion-parameter Kolibri, but recommends less than its one-million-token maximum for complex work

Aleph Alpha released Kolibri on October 3. It published the German-English model’s weights under Apache 2.0. The company built Kolibri for document analysis, reasoning and tool use.

The model has 78.1 billion parameters, but it uses about 3.46 billion for each token. It selects a small part of the model at each step, which cuts the computation needed. The full weights still need substantial memory.

Aleph Alpha advertises a maximum context of 1,048,576 tokens. The model card says the model was trained with a native length of 262,144 tokens. It recommends staying at or below that length for efficient serving and complex tasks. The company says it tested the longer context, but no independent evaluation is cited here.

The published eight-bit weights take up about 78 GB. The minimum hardware list includes two 80 GB A100 GPUs or one H200, B200 or B300. Downloadable weights give organizations another way to deploy the model. They do not eliminate hardware or operating costs.

Aleph Alpha supplied these claims and has an interest in the result. Its benchmark and efficiency comparisons have not been independently verified here. Open weights and an Apache 2.0 licence do not prove that the model gives reliable answers. They also do not prove that it lowers total costs or complies with the rules for a particular application.

Organizations that process German and English documents can test the model on their own infrastructure. That gives them more control over where their data stays and how they deploy the model.

Independent tests should show how Kolibri handles German and English documents, especially when evidence is missing or the task uses a long context. Memory use and the cost of each completed task matter too.